TL;DR
- Voice is constant — tone shifts with context. The Mailchimp distinction is the spine of every working voice document.
- A voice document written for writers is a reference. The same document, with one addendum, becomes operational input for AI-augmented production.
- Specific examples beat abstract description — five to fifteen on-brand paragraphs anchor a prompt better than four adjectives ever did.
- Banned-phrase lists outperform style descriptions. Persona references help voice. They hurt facts. Keep statistics outside the persona frame.
- At fifty pieces a month, voice consistency requires three layers — voice bible as retrieval context, a classifier as second pass, and sampled human review of one piece in ten.
The head of content opens the voice document for the first time in eighteen months. The file is a tidy nine pages. Voice description, four axes with sensible placements, vocabulary preferences, six worked examples that have aged surprisingly well.
Two writers used to read the document on their first day. The team has fourteen contributors now, half of them part-time. A drafting agent produces the first pass on most pieces before any human reads them.
The document is not wrong. The document is also no longer the right shape for the job it now does.
That is the shift this piece is about.
What changed about the brand voice document between 2018 and 2026?
The artifact looks identical. The reader changed.
A 2018 voice document was an onboarding tool. A new writer joined the team, read the document, and produced first drafts that needed light editing. The document succeeded when it shortened the ramp from three months to three weeks.
Voice description, four axes, principles, tone shifts, vocabulary preferences, worked examples. That is the standard six-section anatomy.
Mailchimp’s Content Style Guide popularized the shape. Atlassian Design Writing and Pretty Fly’s Brand Voice Bible methodology refined it. The shape held up for a decade.
A 2026 voice document does both jobs. It still onboards human writers. It also feeds an agent that produces a first draft for nearly every blog post, every email, every landing-page revision, and most social copy.
The agent reads the document differently than a person does. A person reads the description and absorbs the worked examples by example. The agent reads literal text and references it adaptively per draft. Specific paragraphs that an agent can retrieve outweigh general adjectives the agent cannot operationalize.
The shift in reader is the whole story. The artifact’s primary 2026 job is feeding prompts, not training new copywriters. Brands that did the work for one purpose now have the asset for the other. Brands that never did the work face the same gap they faced before, with the cost compounding monthly.
Why does AI default to corporate-speak unless you forbid it?
A drafting agent is a pattern-matching system. It learned from the corpus where business writing actually lives — press releases, product pages, LinkedIn posts.
The corpus is dominated by passive constructions, empty intensifiers, generic adjectives, status-marker jargon, and stock announcement openers. The agent defaults to that pattern because the pattern is the statistical middle of the data the agent learned from.
The default is not a bug. The default is what the agent does when nothing in the prompt forbids it.
Three forces compound the problem in 2026. The training corpus grew faster than the share of well-edited copy inside it. More marketing teams ship more content per week, much of it draft-only-edited.
Some of that copy becomes training data for the next generation of agents. The middle of the distribution gets denser. The agent’s default gets more confident.
The implication for the voice document is direct. A voice description that says "we are warm and direct" gives the agent no signal stronger than the corpus it trained on. A banned-phrase list with concrete entries gives a signal sharp enough to override the default.
The list is operational. The description is decorative. Concrete entries look like innovative, next-level, seamless, cutting-edge, robust, and we’re excited to announce — words to forbid by name.
The cmswire 2026 piece on customer experience names the consequence in plainer language. "The first letter in AI stands for artificial," the piece argues. What customers reach for under stress is "another A: authenticity."
Generic copy reads as artificial because it was averaged. Specific copy reads as authentic because it was edited away from the middle. The voice document’s job is to push the agent off the middle.
What does an AI-prompting addendum to the voice document actually contain?
A new section is emerging in working voice documents — the AI-prompting addendum. The addendum sits at the back of the document. It contains the operational pieces an agent can use directly.
Five components recur across the addendums that work.
Five to fifteen example paragraphs the agent treats as voice anchors. Specific, recent, on-brand. With the small voice tells the writer would recognize.
Search Engine Land’s 2026 guidance puts the band at five to fifteen for prompt-engineering use. Thirty to two hundred or more documents for a retrieval-augmented setup.
A banned-phrase list with rationale. The list names the words and openers the brand has chosen to leave out. The rationale matters because the editor will need to defend the list against contributors who want to add a word back. Twenty to fifty entries is the band most working lists settle on.
Reading-level and sentence-length caps. A reading-level target ("plain English, accessible to a non-technical buyer") and a sentence-length cap ("twenty-five words maximum") measurably shift the output register. Without rewriting the rest of the prompt.
A persona reference for voice — and the explicit instruction not to use it for facts. The Register’s 2024 reporting on persona prompting found a tradeoff. Expert personas raised writing quality and degraded fact accuracy.
The MMLU benchmark dropped from 71.6 percent baseline to 68.0 percent under an expert-persona frame. Split the prompt — persona for voice, raw constraints for any sentence that contains a statistic or a regulated claim.
A fact-handling rule. The rule names where the agent must verify before asserting. Numbers, dates, product names, cited sources. And where the agent may write fluently from the drafting prompt alone — transitions, framing, examples drawn from the worked-examples list.
The addendum is short. Three to five pages of working text. The shortness is part of why it works.
How do voice examples beat voice descriptions for prompting?
Voice description tells the agent what the voice is supposed to be. Voice examples show the agent what the voice actually sounds like.
The mechanism is mundane. An agent has read tens of thousands of brand-voice descriptions. "Warm and direct" appears across brands whose copy is warm and direct. It also appears across brands whose copy is cold and vague.
And brands whose copy is overheated and hedged at once. The phrase has no clean signature in the training data. It cannot pull the agent’s output reliably toward one register.
A specific paragraph the brand has actually shipped has a signature the agent can match. The vocabulary is the brand’s vocabulary. The sentence rhythm is the brand’s rhythm. The relationship to the reader is the brand’s relationship.
The agent can produce a fourth paragraph that fits because three paragraphs are enough to constrain the pattern.
A working voice-examples block has three properties. It is recent (within the past twelve months, ideally). It spans contexts — a long-form blog paragraph, a product-page hero line, an email opening, a social caption. It includes a side-by-side before-after pair.
The before is the kind of generic line the agent would have produced unprompted. The after is what the brand actually shipped.
The before-after pair is the most-load-bearing piece. The pair teaches the agent the edit, not just the destination. The Toolient before-after rules illustrate the pattern. "Lightweight design" becomes "carry it all day without shoulder fatigue." "High-quality materials" becomes "lasts three or more years under daily use." The agent reads the edit and applies the same kind of move on the next sentence.
For more on diagnosing whether a draft needs editing or rewriting from the prompt up, see humanize the draft or write a new prompt.
Why do banned-phrase lists outperform style adjectives?
A banned-phrase list operates at the layer where corporate-speak lives — the actual words on the page. A style description operates one altitude above. The agent has to decide which words satisfy the description.
The asymmetry favors the list. Telling an agent "be concise" is a goal the agent has to satisfy on every sentence. The agent has wide latitude to interpret what concise means.
Telling the agent to drop the cliche cluster is a constraint the agent applies token by token. Status-marker verbs, stock openers, empty intensifiers, named one by one. The agent has no latitude.
A working banned-phrase list has three layers. The universal layer covers the AI-cliche set. The cluster of generic adjectives, status-marker verbs, and stock openers every brand should ban regardless of voice.
The brand-specific layer covers phrases the brand has decided not to use. Words whose connotation drifts away from the brand’s chosen position on the voice axes. The contextual layer covers phrases banned for specific contexts. Words that work in one channel and not in another.
That last layer hints at a bigger question. The same voice has to sound right in an email, on a landing page, and in a short post. Knowing what changes across channels keeps the voice constant while the format moves.
The list is maintained, not authored once. New entries arrive monthly because the AI-cliche set itself drifts. New phrases get over-produced as the corpus updates.
They earn their way onto the list. Old entries occasionally come off because the brand’s voice has moved.
The discipline is procedural. Whoever runs editorial review keeps a running tally of phrases the editor has had to strike from drafts. When a phrase appears in three drafts within a month, the phrase moves to the banned-phrase list.
The next prompt forbids it. The pattern stops.
When does a persona reference help, and when does it hurt the facts?
A persona reference is one line in the prompt that names the role the agent is playing. "You are a senior B2B copywriter writing for a regulated-industries audience." Or "You are the head of content at a plain-language consumer brand." The line is short. It measurably shifts the output’s voice register.
The line also measurably degrades fact accuracy when the prompt contains anything the agent must get right.
The Register’s 2024 reporting is the source most often cited. Across alignment-dependent tasks (writing, role-play, safety) personas improved performance. Across knowledge-dependent tasks (math, coding, fact-fetching), expert personas reduced accuracy.
The MMLU benchmark dropped from 71.6 percent baseline to 68.0 percent under an expert-persona frame. The mechanism is plausible but not fully understood. The directional finding is robust enough to act on.
The implication for voice prompting is a split prompt. The persona frames the voice section of the input. Statistics, product specifications, regulatory claims, named third-party sources, and any sentence the legal team would review live outside the persona frame.
They sit in a separate constraint block. With explicit instructions to write only what the prompt provides and to flag anything not covered.
The split is mechanical. The voice document’s persona reference goes in the addendum. The fact-handling rule sits next to it. Every drafting prompt uses both.
The agent gets the voice signal where voice is what it is producing. The agent gets the constraint signal where accuracy is the work.
What does voice consistency at fifty pieces a month require?
A single editor reading every draft can hold voice across about ten pieces a month. Beyond that volume, voice drift compounds faster than one editor can catch.
The 2026 production-volume tiers settle into a rough pattern. Up to ten pieces a month, prompt-anchored voice and one editor suffices. Ten to fifty pieces, a multi-writer team begins to drift. The upgrade path is retrieval-augmented voice plus a voice classifier as second-pass review.
Fifty to two hundred pieces, the discipline becomes the three-layer pattern. Voice bible indexed as retrieval context. Classifier review on every draft.
Sampled human review of five to ten percent of pieces. Above two hundred pieces, specialist roles emerge — voice-bible maintainer, classifier retrainer, drift detector tracking editorial signal week over week.
The drift signal that matters is editorial revision rate. AirOps’s mechanical observable — the rate at which editors are rewriting the same kinds of voice issues across drafts. The signal is more useful than a static pass-rate score because it tracks change.
A 75 percent pass-rate alone is a snapshot. Editorial revision patterns tell the team whether the snapshot is improving, stable, or eroding. At production volume the trajectory matters more than the snapshot.
The infrastructure is what the voice document becomes. The document is no longer the artifact a writer reads on day one. The document is the corpus a retrieval layer reads continuously.
It is the training set a classifier learns from. It is the reference set a sampled human review compares against. Voice-as-infrastructure is a literal description.
For more on the AI-cliche set the banned-phrase list begins from, see the AI-writing cliche list.
Other questions worth answering
How do tone shifts get captured separately from the constant register?
Mailchimp’s Content Style Guide treats register as constant and tone as the modulation per context. The cleanest pattern for 2026 work is a small tone-shift map — one row per context like apology, launch, status update, support reply. Each row carries two example sentences and the moves the writer must make. The map sits next to the worked examples, not inside them.
Where does the raw vocabulary anchoring a credible written register come from?
From the customers, not the marketing team. Joanna Wiebe’s voice-of-customer method works like this — pull roughly 20 to 50 verbatim quotes out of reviews, support tickets, and post-sale interviews. Cluster the quotes by theme and rank the highest-frequency turns of speech. Drop those exact phrases into headlines, value props, and the worked-examples section your drafting agent reads.
What are the early signs that a multi-writer team is drifting from its codified register?
Three patterns to watch. First, identical paragraphs sound subtly different across channels — one writer’s social posts feel sharper than another’s blog work. Second, drafting agents produce copy that reads more generic than it did 6 months earlier. Third, editors keep rewriting the same kinds of moves on every draft, the AirOps revision-rate diagnostic.
When does a deliberately restricted register serve readers better than the team’s natural one?
When the cost of a misread is high. Government services, healthcare patient education, and financial disclosure pages run on tight constraints — short sentences, common words, active voice, one idea per sentence. GOV.UK pioneered the published version. The discipline is a deliberate choice to maximize comprehension, not a downgrade in sophistication.
Does a tightly codified register correlate with how engines pick up mentions of the company?
Not directly proven, but probably correlated through E-E-A-T and entity-graph signals. The cmswire piece from February 2026, Human First, AI Smart, frames the boundary. The author argues that the first letter in AI stands for artificial, and that customers reach for another A, authenticity. A codified register helps copy land as authentic across owned and third-party surfaces.
Which line of your voice document would you rewrite for AI use first?
Pick one line. The voice description’s third sentence is usually the one that matters.
Most voice documents open with three sentences describing what the brand sounds like. The first two are usually fine — specific enough to carry information. The third sentence is where the document hedges.
It generalizes, it qualifies, it adds a comma-spliced reassurance. The sentence reads like brand-strategy prose because that is what it was. A strategy artifact, not a prompt input.
Rewrite the third sentence as an instruction. Replace "we strive to balance professionalism with approachability" with something concrete. Something like, "we use contractions in body prose, we avoid status-marker verbs, and we end every paragraph on a concrete noun." The first version is description. The second is constraint.
An agent reading the second version produces measurably different output than an agent reading the first.
Read the rewritten line aloud. Does the line name specific words to use or to avoid? Specific rhythms to keep or to break? Specific moves the agent must make on the page?
Then the line is doing the new job. If the line still sounds like brand strategy, the line is still doing the old job.
If your voice document was written before AI-augmented production became routine and the team is feeling the drift, you can contact me here. Send the document and one paragraph of recent output that did not feel right. I will read both and mark the three lines in the document that would do the most work as constraints. There is no charge and no follow-up sales call.